OpenAI has published research finding that about one in six work-related ChatGPT messages asks for a task associated with a different occupation than the user's own job. Once generic activities like writing, summarizing and scheduling are stripped out, that share rises to just under half. The study analyzed more than 800,000 messages from US individual accounts, and it is OpenAI's own research on OpenAI's own product.
Key facts
- 800,000+ work-related messages from US individual ChatGPT accounts, randomly sampled.
- About one in six classified as another occupation's task; just under half after excluding generic work.
- Eight occupation groups retained, with roles taken from self-reported Department and Role metadata.
- Primary source: OpenAI's post How AI is expanding what people do at work and its full report and appendix, titled Work at the Frontier.
What the number actually means
The phrase doing the work here is "classified as." OpenAI used ChatGPT itself to map each message hierarchically onto the O*NET catalogue of work activities - the US Labor Department's standard taxonomy of what jobs consist of - using up to nine preceding messages as context. It then compared that classification against the occupation the user had entered in their account.
So the finding is: a marketer asked for something O*NET files under an analyst's job. A support lead asked for something filed under a developer's. That is genuinely interesting. It is not the same as observing that the marketer did the analyst's work, replaced an analyst, or produced anything usable.
The denominator shift deserves attention too. One in six becomes nearly half only after removing "generic" activities - writing, summarizing, scheduling - that belong to almost every job. That is a defensible methodological choice, and it is also the choice that produces the headline-friendly number. Nearly half of work has not crossed occupational boundaries; nearly half of non-generic prompts were classified outside the user's role.
Why it matters anyway
The underlying phenomenon is real and worth naming. The traditional shape of specialization is that you do not attempt a task outside your training because the cost of starting is too high - you do not know the tools, the conventions or the vocabulary. What a capable general-purpose model changes is the cost of the first attempt. You can now produce a draft SQL query, a draft contract clause, a draft press statement without the years that used to gate entry.
That is a meaningful shift in how work gets distributed inside an organization, and OpenAI is well positioned to observe it because it sees the requests. Nobody else has this data.
The honest caveat
This is vendor research on the value of the vendor's own product, produced by OpenAI Economic Research using OpenAI's proprietary usage window, with OpenAI authors and no disclosed external funder. That does not make it wrong. It does mean the framing choices - which activities count as generic, how boundary-crossing is defined - were made by a party with an interest in the answer.
OpenAI is admirably explicit about the limits. The sample is not representative of the US workforce. Enterprise users are excluded. Only eight role groups are covered. The unit of analysis is a message, not an hour, a project or a job. And critically, the study does not observe whether the output was used, whether it was correct, whether anyone reviewed it, whether it saved time, whether the task was possible without AI, or whether anyone's employment changed.
The missing condition arrived the same day from an unlikely source. Starbucks retired an AI inventory tool from 11,300 cafes after nine months because it miscounted stock. A worker attempting an adjacent task is the beginning of the story; whether the attempt held up in a messy real workflow is the rest of it, and usage data cannot see that part.
For the broader picture, our earlier coverage of Stanford's search for an AI jobs shock and what layoff filings actually attribute to AI covers what the labour-market data does and does not show. The most accurate reading of this study is a credible description of how a selected population of ChatGPT users behaves - not a finding that ChatGPT increases productivity or safely enables generalist work.
Originally published on Ground Truth, where every claim is checked against the primary source.
